SoMark’s launch post for OXR on 20 October 2026. It claims first place on olmOCR-Bench, OmniDocBench and DocParsingBench and first place in parsing speed among open-source models; this piece checks those claims against SoMark’s own benchmark document, which is the source of every figure here.
fixing a document parser on hard tables, formulas, or multi-column layouts usually stalls on manual labeling. OXR (Optical Everything Recognition) is pretrained from scratch and is the first document parser to use a recursive self-improvement (RSI) loop in post-training: OXR parses unlabeled pages, a general VLM and disagreement across OXR rollouts flag errors and uncertain samples, a clustering model groups them by cause, and the VLM plans the next training round and generates targeted training data automatically.
for inference, OXR is also the first to make layout analysis independent and pluggable while running recognition and content-aware reading order in parallel inside the same 1.3B model, so teams can replace the built-in layout model with a custom one for their own document structures. it outputs agent-friendly Markdown or JSON with layout coordinates across 13 element types and ranks #1 on olmOCR-Bench (87.84), OmniDocBench (97.18), and DocParsingBench (87.76), while leading open-source speed on a single RTX PRO 6000 (100 pages in 101.0s, 1.43x PaddleOCR-VL-1.6 and 15.43x Chandra-OCR-2).
— AshutoshShrivastava (@ai_for_success) Oct 23 14:39
A post on 23 October 2026 by the @ai_for_success account summarising how OXR works and repeating SoMark’s scores and speed figures. We use it for context only; the figures in this piece come from SoMark’s GitHub.
SoMark has released OXR-1.0, an open document-parsing model with 1.3 billion parameters, and in its own benchmark tables it ranks first on three tests: 87.84 on olmOCR-bench, 97.18 on OmniDocBench and 87.76 on DocParsingBench. SoMark announced it on X on 20 October 2026. The scores are self-reported: SoMark has asked the OmniDocBench team to re-test them, and that request was still open on 24 October.
The weights are free to download, but the licence is not plain Apache 2.0. The OXR License says commercial use is permitted without a separate commercial licence only while a user and its affiliates do not exceed “USD 2,000,000 in Total Funding” or “USD 2,000,000 in Annual Revenue.” Separately, at any size, it bars using OXR to run a hosted API that competes with SoMark’s own.
SoMark’s website calls the company “Document Intelligence Infrastructure for Agents.” Its launch post said OXR would “Turn documents into Agent-Friendly data.”
Business Pill · OPEN WEIGHTS
A short Business Engineer explainer on open weights: renting a model versus running it yourself. It teaches the general idea only and says nothing about SoMark or OXR.
The key insight: As we read it, the number to repeat is the licence line, not only the scores: SoMark puts a 1.3B model that ranks first in its own tables into anyone’s hands, and asks for a separate commercial licence once a user’s funding or revenue exceeds $2 million.
What SoMark Released
The model. The GitHub README describes OXR as “An open-source document parsing model with SOTA accuracy and speed, using an RSI loop during post-training.” It says OXR was pretrained from scratch, and that the recursive self-improvement (RSI) loop is “a first in document parsing.”
What it reads. OXR recognises 13 element types, according to the README, including tables, formulas, chemical equations, chemical structures and code blocks, with bounding-box coordinates. The README says it has “a particularly wide lead in table parsing.”
How it is built. The system is two models: a layout-analysis model and OXR itself, which handles both recognition and reading order. SoMark says the layout model can be swapped out: “OXR is the first to make layout analysis independent and pluggable.”
How to run it. The weights are on Hugging Face and ModelScope in BF16. Running it locally needs Linux and an NVIDIA GPU, with 24GB of GPU memory to start and 48GB or more recommended. SoMark also runs a free API with no key, and warns: “We’ll keep the free API online until we can no longer afford to run it.”
The commercial link. The README says OXR’s API is compatible with SoMark’s official document-parsing API, so SoMark’s own command-line tools and SDKs work with OXR by changing the base URL.
OXR-1.0’s overall scores on three document-parsing benchmarks, each first in SoMark’s own tables, and the $2 million commercial threshold in its licence. The scores are self-reported; a re-test request to OmniDocBench was open on 24 October 2026. Sources: SoMark, OXR README, docs/benchmark.md and LICENSE on GitHub; OmniDocBench issue #264.
The Scores, and Who Has Checked Them
The table gives OXR’s overall score on each benchmark, the next model in SoMark’s own ranking, and the gap between them by our arithmetic.
Benchmark
OXR-1.0
Next in SoMark’s table
Gap (our arithmetic)
olmOCR-bench (17 models)
87.84
Infinity-Parser2-Pro, 87.60
0.24 points
OmniDocBench (23 models)
97.18
TeleOCR, 96.91
0.27 points
DocParsingBench (17 models)
87.76
PaddleOCR-VL-1.6, 85.17
2.59 points
Source: SoMark, docs/benchmark.md on GitHub, read 24 October 2026. SoMark dates the accuracy snapshot 5 October 2026 and writes that “Rankings apply to the displayed models and recorded evaluation protocols.”
Two of the three leads are narrow. On olmOCR-bench and OmniDocBench the gap to the next model is under three-tenths of a point, by our arithmetic. The wider lead is on DocParsingBench. On OmniDocBench’s Table TEDS score, TeleOCR’s 96.82 is above OXR’s 96.62 in SoMark’s own table; on the structure-only TEDS-S score, OXR leads, 98.44 to 98.18.
Where the rival scores come from. SoMark says peer scores on olmOCR-bench are “sourced from published papers, model cards, and the olmOCR-bench leaderboard,” and on OmniDocBench mostly from the benchmark’s own published table. OXR’s scores are SoMark’s own runs.
The outside check. On 19 October a GitHub account writing as OXR’s developers opened an issue on OmniDocBench’s GitHub asking the team to re-test OXR’s self-tested results and add them to the leaderboard. A reply on 21 October said, in Chinese, that the team would confirm after re-testing. In the olmOCR-bench results table on AllenAI’s olmocr GitHub page, read on 24 October, OXR is not listed. SoMark’s benchmark document does not say who maintains DocParsingBench.
The Speed Claim
SoMark timed 21 models on the same 100 OmniDocBench pages on one NVIDIA RTX PRO 6000 Blackwell GPU with 96 GB of memory. It reports that OXR parsed the 100 pages in 100.972 seconds on average across two rounds, with “33.5% higher throughput” than the next model, OvisOCR2.
End-to-end pages per second on the same 100 OmniDocBench pages, one NVIDIA RTX PRO 6000 GPU, as reported by SoMark: the top 8 of 21 models plus Chandra-OCR-2. Self-reported and not independently verified. Source: SoMark, OXR docs/benchmark.md, Table 5.
By our arithmetic from SoMark’s Table 5, that is 1.43 times the page rate of PaddleOCR-VL-1.6 (144.728 seconds) and 15.43 times that of Chandra-OCR-2 (1,557.821 seconds). On raw decoding, SoMark gives OXR 557.77 tokens per second, 8.0% above PaddleOCR-VL-1.6. SoMark notes that native tokens vary by tokenizer.
What the Licence Allows
The licence file opens: “OXR is licensed under the Apache License, Version 2.0, subject to the additional terms below.” Three terms matter for business use.
The $2 million thresholds. Commercial use needs no separate licence while a user and its affiliates, on a consolidated basis, exceed neither $2 million in total funding nor $2 million in annual revenue. If either is exceeded, the licence says a user “must obtain a separate commercial license from SoMark Team before commencing or continuing commercial use.”
No competing hosted API. Without written permission, users may not offer third parties “a hosted API service that competes with a hosted API service offered by SoMark Team or its Affiliates.” This covers document parsing, OCR, layout analysis and content extraction, and the licence says it “applies regardless of Total Funding or Annual Revenue.” Use inside a user’s own organisation is not barred by that section.
Attribution and termination. An online service built on OXR must “prominently identify SoMark OXR” in its interface or public documentation. Rights end automatically if a user breaks any of the three terms.
Connecting the Coverage
OXR is the latest small, specialised model we have covered this week. Our Meta tool-selection piece reported a paper by Meta and Carnegie Mellon researchers in which an 8-billion-parameter open model, trained for one narrow job, picked the exact right tools for 41.5% of test requests, against 30.6% for the 402-billion-parameter Llama-4-Maverick.
Our Perplexity piece covered a 4B decision model that Perplexity now lets users run on a Mac with no API key, while the same model lists at $0.01 per million input tokens on its hosted API. Our Abacus piece covered Smaug-Nemotron-3.5, an open-weight agent fine-tune that costs $0.197 per successful task on Abacus’s own leaderboard, which, like OXR’s tables, is run by the company whose model it ranks.
On licence terms, our open-versus-fenced piece counted four open AI launches and four new controls in nine days, and reported that MiniMax’s licence asks commercial products generating more than $20 million in yearly revenue for prior written authorization. OXR’s commercial threshold is a tenth of that, at $2 million, and it also counts funding.
The Structural Read
As we read it, document parsing is the data layer under agents and retrieval systems. SoMark pitches OXR as a way to “Turn documents into Agent-Friendly data”, and the README says it outputs Markdown or JSON with layout coordinates, the form an agent or a retrieval index reads.
As we read it, that is the pressure point for paid OCR and document-parsing APIs: a model SoMark ranks first runs on one GPU with 24GB of memory to start, per the README, and SoMark also offers a free API with no key, which it says is not for production.
As we read it, the licence is how SoMark keeps its paid layer. The weights are open, but the licence says no one may use OXR to run a competing hosted parsing API, at any size, and users above $2 million in funding or revenue need a separate licence. MiniMax’s licence, in our earlier coverage, drew its line at $20 million in yearly revenue.
OXR License, Section 3
“This restriction applies regardless of Total Funding or Annual Revenue.”
Three Implications
FOR TEAMS BUILDING AGENTS AND RAG SoMark’s README says OXR turns PDFs and document images into Markdown or JSON with layout coordinates across 13 element types; the free API lets teams test it, though SoMark says it does not recommend it for production.
FOR PAID PARSING APIS The licence says no one may use OXR to offer third parties a hosted parsing, OCR, layout-analysis or extraction API that competes with SoMark’s, whether paid or free, without SoMark’s written permission.
FOR BUYERS READING BENCHMARKS SoMark’s leads on olmOCR-bench and OmniDocBench are 0.24 and 0.27 points by our arithmetic, and OmniDocBench’s own re-test was pending on 24 October.
The framework states the pattern as “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, OXR follows that split: SoMark opens the model weights, while its licence keeps the hosted API, the layer SoMark sells, closed to competitors. The framework was published in March 2026 and does not discuss SoMark or OXR.
What Is Not Established
No outside party has yet published OXR scores in the material we read: the OmniDocBench re-test was pending on 24 October, and the olmOCR table we read does not list OXR. SoMark’s pages that we read do not give the price of a separate commercial licence, the company’s funding or revenue, or where it is based. There is no technical report yet: the README says “We will add a paper citation once the technical report is published.”
SoMark’s OXR-1.0 is a 1.3-billion-parameter open document parser that ranks first on olmOCR-bench, OmniDocBench and DocParsingBench in SoMark’s own tables, and fastest of 21 models in SoMark’s own speed test. The leads on two of the three benchmarks are under three-tenths of a point, and an OmniDocBench re-test is pending. The weights are free to download; commercial use above $2 million in funding or revenueneeds a separate licence, and no one may use OXR to run a competing hosted parsing API.
A note on sourcing. This piece rests on SoMark’s GitHub README, benchmark document and licence for OXR, SoMark’s launch post on X, OmniDocBench’s public GitHub issue tracker, AllenAI’s olmocr results table and our own earlier pieces, all read on 24 October 2026. Every benchmark and speed figure is SoMark’s own and has not been independently verified; comparisons and gaps marked as our arithmetic are calculated from SoMark’s tables. We did not contact SoMark.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
